File size: 11,005 Bytes
9b7cf70 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 | #!/usr/bin/env python3
"""
DI LeRobot Converter API
========================
Receives episode data (JSON + video URL) from the iOS app,
creates a LeRobot v2.0 parquet file, uploads parquet + video
to the HuggingFace dataset repo, and updates meta/info.json.
Deployed as a HuggingFace Space with Gradio.
The iOS app calls the /api/convert endpoint after uploading to GCS.
"""
import gradio as gr
import json
import os
import tempfile
import shutil
from pathlib import Path
import pandas as pd
import numpy as np
from huggingface_hub import HfApi, hf_hub_download
# Config
HF_DATASET_REPO = "DynamicIntelligence/humanoid-robots-training-dataset"
HF_TOKEN = os.environ.get("HF_TOKEN", "")
GCS_BUCKET = "di_record_intern_data"
CHUNKS_SIZE = 100
def convert_episode(episode_json: str) -> str:
"""
Convert episode data to LeRobot v2.0 format and upload to dataset repo.
Input JSON schema:
{
"episode_index": int, # auto-assigned if -1
"language_instruction": str,
"fps": int,
"frames": [
{
"timestamp": float,
"pose": {"x": f, "y": f, "z": f, "yaw": f, "pitch": f, "roll": f},
"left_hand": [x, y, z] or null,
"right_hand": [x, y, z] or null
}, ...
],
"video_gcs_path": str # GCS path to rgb_video.mp4
}
"""
try:
data = json.loads(episode_json)
except json.JSONDecodeError as e:
return json.dumps({"error": f"Invalid JSON: {e}"})
api = HfApi(token=HF_TOKEN)
# Determine episode index
episode_index = data.get("episode_index", -1)
if episode_index < 0:
# Auto-assign: read current info.json to get next index
try:
info_path = hf_hub_download(
repo_id=HF_DATASET_REPO, filename="meta/info.json",
repo_type="dataset", token=HF_TOKEN
)
with open(info_path) as f:
info = json.load(f)
episode_index = info.get("total_episodes", 0)
except Exception:
episode_index = 0
lang = data.get("language_instruction", "")
fps = data.get("fps", 30) or 30
frames = data.get("frames", [])
num_frames = len(frames)
if num_frames == 0:
return json.dumps({"error": "No frames in episode data"})
# Build parquet rows
rows = []
for i, frame in enumerate(frames):
pose = frame.get("pose", {})
cam_x = pose.get("x", 0)
cam_y = pose.get("y", 0)
cam_z = pose.get("z", 0)
cam_roll = pose.get("roll", 0)
cam_pitch = pose.get("pitch", 0)
cam_yaw = pose.get("yaw", 0)
camera_pose = [cam_x, cam_y, cam_z, cam_roll, cam_pitch, cam_yaw]
# Hand data: [x, y, z] from end_effector → pad to 9 values (3 joints × xyz)
lh = frame.get("left_hand") or [0, 0, 0]
rh = frame.get("right_hand") or [0, 0, 0]
# Pad single palm position to 3-joint format (wrist=palm, others=0)
left_hand = list(lh[:3]) + [0.0] * 6
right_hand = list(rh[:3]) + [0.0] * 6
# Action deltas
if i > 0:
prev = frames[i - 1]
pp = prev.get("pose", {})
prev_cam = [pp.get("x", 0), pp.get("y", 0), pp.get("z", 0),
pp.get("roll", 0), pp.get("pitch", 0), pp.get("yaw", 0)]
cam_delta = [camera_pose[j] - prev_cam[j] for j in range(6)]
plh = prev.get("left_hand") or [0, 0, 0]
prh = prev.get("right_hand") or [0, 0, 0]
lh_delta = [lh[j] - plh[j] if j < len(lh) and j < len(plh) else 0 for j in range(3)] + [0.0] * 6
rh_delta = [rh[j] - prh[j] if j < len(rh) and j < len(prh) else 0 for j in range(3)] + [0.0] * 6
else:
cam_delta = [0.0] * 6
lh_delta = [0.0] * 9
rh_delta = [0.0] * 9
rows.append({
"episode_index": episode_index,
"frame_index": i,
"timestamp": frame.get("timestamp", i / fps),
"observation.camera_pose": camera_pose,
"observation.left_hand": left_hand,
"observation.right_hand": right_hand,
"action.camera_delta": cam_delta,
"action.left_hand_delta": lh_delta,
"action.right_hand_delta": rh_delta,
"language_instruction": lang,
"next.done": i == num_frames - 1,
})
# Create parquet
tmp = Path(tempfile.mkdtemp())
try:
df = pd.DataFrame(rows)
chunk_idx = episode_index // CHUNKS_SIZE
parquet_path = tmp / f"episode_{episode_index:06d}.parquet"
df.to_parquet(parquet_path, index=False)
# Upload parquet
api.upload_file(
path_or_fileobj=str(parquet_path),
path_in_repo=f"data/chunk-{chunk_idx:03d}/episode_{episode_index:06d}.parquet",
repo_id=HF_DATASET_REPO, repo_type="dataset", token=HF_TOKEN,
)
# Upload video from GCS if provided
video_gcs_path = data.get("video_gcs_path", "")
video_gcs_url = data.get("video_gcs_url", "")
video_uploaded = False
if video_gcs_url:
# Download from GCS public URL and re-upload to HF
import urllib.request
video_local = tmp / "rgb_video.mp4"
try:
urllib.request.urlretrieve(video_gcs_url, str(video_local))
api.upload_file(
path_or_fileobj=str(video_local),
path_in_repo=f"videos/chunk-{chunk_idx:03d}/rgb/episode_{episode_index:06d}.mp4",
repo_id=HF_DATASET_REPO, repo_type="dataset", token=HF_TOKEN,
)
video_uploaded = True
except Exception as ve:
pass # Video upload is optional
# Update meta/info.json
try:
existing_info_path = hf_hub_download(
repo_id=HF_DATASET_REPO, filename="meta/info.json",
repo_type="dataset", token=HF_TOKEN
)
with open(existing_info_path) as f:
info = json.load(f)
info["total_episodes"] = max(info.get("total_episodes", 0), episode_index + 1)
info["total_frames"] = info.get("total_frames", 0) + num_frames
info["splits"] = {"train": f"0:{info['total_episodes']}"}
info["total_chunks"] = (info["total_episodes"] - 1) // CHUNKS_SIZE + 1
if video_uploaded:
info["total_videos"] = info.get("total_videos", 0) + 1
except Exception:
info = build_default_info(episode_index, num_frames)
meta_dir = tmp / "meta"
meta_dir.mkdir(exist_ok=True)
with open(meta_dir / "info.json", "w") as f:
json.dump(info, f, indent=2)
api.upload_folder(
folder_path=str(meta_dir), path_in_repo="meta",
repo_id=HF_DATASET_REPO, repo_type="dataset", token=HF_TOKEN,
)
result = {
"success": True,
"episode_index": episode_index,
"num_frames": num_frames,
"parquet_uploaded": True,
"video_uploaded": video_uploaded,
"dataset_url": f"https://huggingface.co/datasets/{HF_DATASET_REPO}",
}
return json.dumps(result)
finally:
shutil.rmtree(tmp, ignore_errors=True)
def build_default_info(episode_index, num_frames):
return {
"codebase_version": "v2.0",
"robot_type": "unknown",
"total_episodes": episode_index + 1,
"total_frames": num_frames,
"total_tasks": 1,
"total_videos": 1,
"total_chunks": 1,
"chunks_size": CHUNKS_SIZE,
"fps": 30,
"splits": {"train": f"0:{episode_index + 1}"},
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
"video_path": "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4",
"features": {
"observation.camera_pose": {"dtype": "float32", "shape": [6],
"names": ["x", "y", "z", "roll", "pitch", "yaw"]},
"observation.left_hand": {"dtype": "float32", "shape": [9],
"names": ["wrist_x", "wrist_y", "wrist_z", "thumb_x", "thumb_y", "thumb_z",
"index_x", "index_y", "index_z"]},
"observation.right_hand": {"dtype": "float32", "shape": [9],
"names": ["wrist_x", "wrist_y", "wrist_z", "index_x", "index_y", "index_z",
"middle_x", "middle_y", "middle_z"]},
"action.camera_delta": {"dtype": "float32", "shape": [6],
"names": ["dx", "dy", "dz", "droll", "dpitch", "dyaw"]},
"action.left_hand_delta": {"dtype": "float32", "shape": [9],
"names": ["wrist_dx", "wrist_dy", "wrist_dz", "thumb_dx", "thumb_dy",
"thumb_dz", "index_dx", "index_dy", "index_dz"]},
"action.right_hand_delta": {"dtype": "float32", "shape": [9],
"names": ["wrist_dx", "wrist_dy", "wrist_dz", "index_dx", "index_dy",
"index_dz", "middle_dx", "middle_dy", "middle_dz"]},
"language_instruction": {"dtype": "string", "shape": [1], "names": None},
"timestamp": {"dtype": "float64", "shape": [1], "names": None},
"frame_index": {"dtype": "int64", "shape": [1], "names": None},
"episode_index": {"dtype": "int64", "shape": [1], "names": None},
"next.done": {"dtype": "bool", "shape": [1], "names": None},
"rgb": {"dtype": "video", "shape": [480, 640, 3],
"names": ["height", "width", "channels"],
"video_info": {"video.fps": 30, "video.codec": "h264",
"video.pix_fmt": "yuv420p", "video.is_depth_map": False,
"has_audio": False}},
},
"videos": {
"rgb": {"video_info": {"video.fps": 30, "video.codec": "h264",
"video.pix_fmt": "yuv420p", "video.is_depth_map": False,
"has_audio": False}}
},
}
# Gradio UI (also exposes /api/convert endpoint automatically)
demo = gr.Interface(
fn=convert_episode,
inputs=gr.Textbox(label="Episode JSON", lines=10, placeholder="Paste episode JSON here..."),
outputs=gr.Textbox(label="Result"),
title="DI LeRobot Converter",
description="Converts episode data from DI iOS app to LeRobot v2.0 format and uploads to HuggingFace dataset repo.",
api_name="convert",
)
if __name__ == "__main__":
demo.launch(server_name="0.0.0.0", server_port=7860)
|